前往 ai.google.dev 查看 | 在 Google Colab 中執行 | 在 Vertex AI 中開啟 | 前往 GitHub 查看原始碼 |
Gemma 等大型語言模型 (LLM) 可生成資訊豐富的回覆,非常適合用來建構虛擬助理和聊天機器人。
傳統上,LLM 以無狀態的方式運作,意即缺乏儲存過往對話的固有記憶體。系統會單獨處理每個提示或問題,並忽略先前的互動。不過,自然對話的關鍵在於保留先前互動的背景資訊。為了克服這項限制並讓 LLM 維護對話脈絡,每次向 LLM 發出新提示時,必須明確提供相關資訊,例如對話記錄或相關部分。
本教學課程說明如何使用 Gemma 經過指令調整的模型變化版本,來開發聊天機器人。
設定
Gemma 設定
如要完成本教學課程,您必須先前往 Gemma 設定頁面完成設定。Gemma 設定操作說明會說明如何執行下列操作:
- 前往 kaggle.com 存取 Gemma。
- 請選取具有足夠資源來執行 Gemma 2B 模型的 Colab 執行階段。
- 產生並設定 Kaggle 使用者名稱和 API 金鑰。
完成 Gemma 設定後,請繼續前往下一節,設定 Colab 環境的環境變數。
設定環境變數
設定 KAGGLE_USERNAME
和 KAGGLE_KEY
的環境變數。
import os
from google.colab import userdata
# Note: `userdata.get` is a Colab API. If you're not using Colab, set the env
# vars as appropriate for your system.
os.environ["KAGGLE_USERNAME"] = userdata.get('KAGGLE_USERNAME')
os.environ["KAGGLE_KEY"] = userdata.get('KAGGLE_KEY')
安裝依附元件
安裝 Keras 和 KerasNLP。
# Install Keras 3 last. See https://keras.io/getting_started/ for more details.
pip install -q tensorflow-cpu
pip install -q -U keras-nlp tensorflow-hub
pip install -q -U keras>=3
pip install -q -U tensorflow-text
選取後端
Keras 是高階的多架構深度學習 API,專為簡化使用而設計。Keras 3 可讓您選擇後端:TensorFlow、JAX 或 PyTorch。這三者都能進行本教學課程。
import os
# Select JAX as the backend
os.environ["KERAS_BACKEND"] = "jax"
# Pre-allocate 100% of TPU memory to minimize memory fragmentation
os.environ["XLA_PYTHON_CLIENT_MEM_FRACTION"] = "1.0"
匯入套件
匯入 Keras 和 KerasNLP。
import keras
import keras_nlp
# for reproducibility
keras.utils.set_random_seed(42)
將模型例項化
KerasNLP 可實作許多熱門的模型架構。在這個教學課程中,您會使用 GemmaCausalLM
將模型例項化,這是用於因果語言模型的端對端 Gemma 模型。因果語言模型會根據先前的符記預測下一個符記。
使用 from_preset
方法將模型例項化:
gemma_lm = keras_nlp.models.GemmaCausalLM.from_preset("gemma_1.1_instruct_2b_en")
Attaching 'metadata.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'metadata.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'task.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'config.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'metadata.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'metadata.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'config.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'config.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'model.weights.h5' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'metadata.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'metadata.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'preprocessor.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'tokenizer.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'tokenizer.json' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook... Attaching 'assets/tokenizer/vocabulary.spm' from model 'keras/gemma/keras/gemma_1.1_instruct_2b_en/3' to your Colab notebook...
GemmaCausalLM.from_preset()
函式會根據預設架構和權重將模型例項化。在上述程式碼中,字串 "gemma_1.1_instruct_2b_en"
指定具有 20 億個參數的預設 Gemma 2B 模型。您也可以使用具有 7B、9B 和 27B 參數的 Gemma 模型。您可以在 Kaggle 的「模型變化版本」清單中,找到 Gemma 模型的程式碼字串。
使用 summary
方法取得模型的詳細資訊:
gemma_lm.summary()
如摘要所示,模型有 25 億個可訓練參數。
定義格式設定輔助函式
from IPython.display import Markdown
import textwrap
def display_chat(prompt, text):
formatted_prompt = "<font size='+1' color='brown'>🙋♂️<blockquote>" + prompt + "</blockquote></font>"
text = text.replace('•', ' *')
text = textwrap.indent(text, '> ', predicate=lambda _: True)
formatted_text = "<font size='+1' color='teal'>🤖\n\n" + text + "\n</font>"
return Markdown(formatted_prompt+formatted_text)
def to_markdown(text):
text = text.replace('•', ' *')
return Markdown(textwrap.indent(text, '> ', predicate=lambda _: True))
建構聊天機器人
Gemma 指令調整模型 gemma_1.1_instruct_2b_en
經過微調,可解讀下列回合符記:
<start_of_turn>user\n ... <end_of_turn>\n
<start_of_turn>model\n ... <end_of_turn>\n
本教學課程會使用這些符記建構聊天機器人。如要進一步瞭解 Gemma 控制權杖,請參閱格式設定和系統操作說明。
建立即時通訊小幫手來管理對話狀態
class ChatState():
"""
Manages the conversation history for a turn-based chatbot
Follows the turn-based conversation guidelines for the Gemma family of models
documented at https://ai.google.dev/gemma/docs/formatting
"""
__START_TURN_USER__ = "<start_of_turn>user\n"
__START_TURN_MODEL__ = "<start_of_turn>model\n"
__END_TURN__ = "<end_of_turn>\n"
def __init__(self, model, system=""):
"""
Initializes the chat state.
Args:
model: The language model to use for generating responses.
system: (Optional) System instructions or bot description.
"""
self.model = model
self.system = system
self.history = []
def add_to_history_as_user(self, message):
"""
Adds a user message to the history with start/end turn markers.
"""
self.history.append(self.__START_TURN_USER__ + message + self.__END_TURN__)
def add_to_history_as_model(self, message):
"""
Adds a model response to the history with start/end turn markers.
"""
self.history.append(self.__START_TURN_MODEL__ + message + self.__END_TURN__)
def get_history(self):
"""
Returns the entire chat history as a single string.
"""
return "".join([*self.history])
def get_full_prompt(self):
"""
Builds the prompt for the language model, including history and system description.
"""
prompt = self.get_history() + self.__START_TURN_MODEL__
if len(self.system)>0:
prompt = self.system + "\n" + prompt
return prompt
def send_message(self, message):
"""
Handles sending a user message and getting a model response.
Args:
message: The user's message.
Returns:
The model's response.
"""
self.add_to_history_as_user(message)
prompt = self.get_full_prompt()
response = self.model.generate(prompt, max_length=1024)
result = response.replace(prompt, "") # Extract only the new response
self.add_to_history_as_model(result)
return result
與模型對話
開始與模型聊天。
chat = ChatState(gemma_lm)
message = "Tell me, in a few words, how to compute all prime numbers up to 1000?"
display_chat(message, chat.send_message(message))
🙋♂️
請詳細回答看看,如何將所有質數計算至 1000?🤖?
Eratosthenes 的圍爐是廣泛使用的方法,可在指定上限內計算所有質數。也就是反覆標示每個質數的倍數。
延續對話,
message = "Now in Python! No numpy, please!"
display_chat(message, chat.send_message(message))
🙋♂️
現已支援 Python!沒騙你,拜託!🤖?
def prime(n): if n <= 1: return False for i in range(2, int(n**0.5) + 1): if n % i == 0: return False return True
message = "Thank you, it works! Can you explain the code in plain English?"
display_chat(message, chat.send_message(message))
🙋♂️
謝謝,您的成功!你能用簡單的英文說明程式碼嗎?🤖?
提供的程式碼會定義一個
prime
函式,用於檢查指定數字是否為質數。運作方式:
這個函式會使用單一引數
n
,也就是要檢查的數字。會先檢查
n
是否小於或等於 1。如果會,則數字不是質數,因此函式會傳回False
。接著進入迴圈,也就是從 2 到
n
的平方根。針對每個數字
i
,它會檢查n
是否以i
平均 (即n % i == 0
).如果
n
被i
整除,則函式會傳回False
,因為如果n
有除數,則無法為基準值。如果迴圈完成,但沒有找到
n
的任何除數,函式會傳回True
,表示n
是質數。使用範例:
>>> prime(2) True >>> prime(3) True >>> prime(4) False >>> prime(5) True
使用本守則的好處:
- 這是一種簡單又有效率的演算法,可找出質數。
- 這項技術廣泛用於各種電腦科學和數學應用程式。
message = "Great! Now add those explanations as comments in the code."
display_chat(message, chat.send_message(message))
🙋♂️
太好了!接著在程式碼中以註解的形式加入說明。🤖?
def prime(n): """ Checks whether a given number is prime or not. Args: n: The number to check. Returns: True if n is prime, False otherwise. """ # Check if n is less than or equal to 1. if n <= 1: return False # Iterate through numbers from 2 to the square root of n. for i in range(2, int(n**0.5) + 1): # Check if n is divisible by i. if n % i == 0: return False # If the loop completes without finding any divisors for n, then n is prime. return True
執行產生的程式碼來測試產生的回應:
def is_prime(n):
"""
Checks if a number is prime.
Args:
n: The number to check.
Returns:
True if n is prime, False otherwise.
"""
# If n is less than or equal to 1, it is not prime.
if n <= 1:
return False
# Iterate through all the numbers from 2 to the square root of n.
for i in range(2, int(n**0.5) + 1):
# If n is divisible by any of the numbers in the range from 2 to the square root of n, it is not prime.
if n % i == 0:
return False
# If no divisors are found, n is prime.
return True
# Initialize an empty list to store prime numbers.
primes = []
# Iterate through all the numbers from 2 to 1000.
for i in range(2, 1001):
# If the number is prime, add it to the list.
if is_prime(i):
primes.append(i)
# Print the prime numbers.
print(primes)
[2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97, 101, 103, 107, 109, 113, 127, 131, 137, 139, 149, 151, 157, 163, 167, 173, 179, 181, 191, 193, 197, 199, 211, 223, 227, 229, 233, 239, 241, 251, 257, 263, 269, 271, 277, 281, 283, 293, 307, 311, 313, 317, 331, 337, 347, 349, 353, 359, 367, 373, 379, 383, 389, 397, 401, 409, 419, 421, 431, 433, 439, 443, 449, 457, 461, 463, 467, 479, 487, 491, 499, 503, 509, 521, 523, 541, 547, 557, 563, 569, 571, 577, 587, 593, 599, 601, 607, 613, 617, 619, 631, 641, 643, 647, 653, 659, 661, 673, 677, 683, 691, 701, 709, 719, 727, 733, 739, 743, 751, 757, 761, 769, 773, 787, 797, 809, 811, 821, 823, 827, 829, 839, 853, 857, 859, 863, 877, 881, 883, 887, 907, 911, 919, 929, 937, 941, 947, 953, 967, 971, 977, 983, 991, 997]
使用 get_history
方法即可查看 Chat
類別如何保留所有背景資訊。
print(chat.get_history())
<start_of_turn>user Tell me, in a few words, how to compute all prime numbers up to 1000?<end_of_turn> <start_of_turn>model The Sieve of Eratosthenes is a widely used method to compute all prime numbers up to a given limit. It involves iteratively marking out multiples of each prime number.<end_of_turn> <start_of_turn>user Now in Python! No numpy, please!<end_of_turn> <start_of_turn>model ```python def prime(n): if n <= 1: return False for i in range(2, int(n**0.5) + 1): if n % i == 0: return False return True ```<end_of_turn> <start_of_turn>user Thank you, it works! Can you explain the code in plain English?<end_of_turn> <start_of_turn>model The provided code defines a function `prime` that checks whether a given number is prime or not. **How it works:** - The function takes a single argument, `n`, which is the number to check. - It first checks if `n` is less than or equal to 1. If it is, the number is not prime, so the function returns `False`. - It then enters a loop that iterates through numbers from 2 to the square root of `n`. - For each number `i`, it checks if `n` is divisible evenly by `i` (i.e., `n % i == 0`). - If `n` is divisible by `i`, the function returns `False` because `n` cannot be prime if it has a divisor. - If the loop completes without finding any divisors for `n`, the function returns `True`, indicating that `n` is a prime number. **Example Usage:** ```python >>> prime(2) True >>> prime(3) True >>> prime(4) False >>> prime(5) True ``` **Benefits of this Code:** - It is a simple and efficient algorithm for finding prime numbers. - It is widely used in various computer science and mathematical applications.<end_of_turn> <start_of_turn>user Great! Now add those explanations as comments in the code.<end_of_turn> <start_of_turn>model ```python def prime(n): """ Checks whether a given number is prime or not. Args: n: The number to check. Returns: True if n is prime, False otherwise. """ # Check if n is less than or equal to 1. if n <= 1: return False # Iterate through numbers from 2 to the square root of n. for i in range(2, int(n**0.5) + 1): # Check if n is divisible by i. if n % i == 0: return False # If the loop completes without finding any divisors for n, then n is prime. return True ```<end_of_turn>
摘要與延伸閱讀
在這個教學課程中,您學到瞭如何在 JAX 上使用 Keras 與 Gemma 2B 指令調整模型進行即時通訊。
請參閱下列指南與教學課程,進一步瞭解 Gemma: